AmpCode is cool

Adding a new feature requires careful observation and accurate perception of the model’s capability and how it reflects in human workflow. Removing a feature also requires the careful insights. Among all the harness and agentic coding tools I’ve used, Amp is the one that does this best, thus it has earned tremendous respect from me.

I first noticed Amp in Lenny’s bundle in early 2026, then started using it on and off, and joined it’s Builder Crew to learn more about it. Now that I’ve been using AmpCode for a while now, I found myself constantly surprised by the interesting features they came up with, so i want to talk about it.

(And also it seems like if i don’t talk about amp then X is not gonna give my posts viewers 😅)

The interesting features they added and removed#

Adding a new feature requires careful observation and accurate perception of the model’s capability and how it reflects in human workflow. Removing a feature also requires the careful insights.

The first interesting feature i discovered was a feature called Thread Map. From my experience working at Heptabase, I understand how diagrams can massively accelerate understanding and relationship tracking. Seeing how if you can weave all the atomic sessions together into a graph sounds interesting, and could potentially reveal some insights of how you work. But in the end I’m not sure what’s the value proposition of this feature, and didn’t use it much. I guess that’s also what the AmpCode team thought so they stopped supporting it now.

Associated with that features is the handoff concept/features. Back then models suffers greatly from context rot, it was a best practice to open a new thread if you want to work on a new task. handoff was designed to help one session drop relevant context to a new session. It was deprecated later this year because of how good compaction had become.

But it didn’t disappear entirely. Handoff has evolved into agents sending messages back and forth between threads, which is incredibly useful. I rely on it frequently when remote Orbs lack sufficient information: I can then simply ask them to send requests to local Amp threads, gather the necessary details, and send the results back. When it can operate across local and remote environments, it offers a fundamental capability that Codex and Claude Code lack.

This brings us to the concept of Orbs. How is Orb different than other remote environments like Codex remote and Claude Code remote? It’s because it offers the same experience as any local environment, you have a browser (portal), terminal, diff views, file tree, and a button that let’s you sync the changes to your local machine. If coding agents are so powerful, why do you need all that? The reason is that not all tasks can be one-shot. There are tasks that you have to iteratively improve, and that has become where engineers spend 80% of their coding time at. And once you need to iteratively improve it, simple chat interface like Codex remote are not enough, as they can only help with the simplest changes.

What’s even better is that Orb unifies local and remote environment controls, allowing you to control anything from any device. It made my iPad useful again after I graduated from college, giving me everything I need to build my app on it.

Now, if you can instantly spin up numerous remote Orbs and run many coding threads, wouldn’t you want a manager to handle all the threads for you? That’s why they introduced Puck. Puck is an orchestration agent that manages all your threads across Orbs. If you integrate it with Slack, it can delegate tasks, spin up machines when needed, and report back to you without losing track of anything.

The No. 1 reason people stopped using Amp was that it was difficult for consumers to use at API rates. With OpenAI’s implicit approval for third-party apps to use ChatGPT subscriptions, Amp introduced an integration that allows users to spend their ChatGPT usage quota in Amp, solving the problem. Amp now has a consumer-friendly pricing model: $20 gives you access to all remote Orbs and includes a limited amount of token usage, while significant additional usage is charged against your own ChatGPT quota.

These descriptions don’t mention that they route different models to specialized tasks. For example, they use the most balanced model, Search, for context retrieval; provide the execution agent with a more capable model, Oracle, for consultation; and use the best image-generation model, called Painter, among others. All of these models are integrated to optimize the balance between cost and performance in a predictable way. And in contrast to model router systems such like the latest Droid router, these are more like subagents the main agent spawns.

Additionally to all these good stuffs, they are serving a much smaller set of people than Codex, Claude Code today, making them listen and iterate a lot faster.

End#

It’s no surprise that Amp is the best system co-evolving with frontier AI capabilities to empower software builders. This is very inspiring, since coding isn’t the only task which evolves with the model’s capability. Every industry should co-evolve. There are infinite work to be done and undone, infinite opportunities to contribute to this great motion.

*They also have a very forward looking manifesto that i will share here.

That above observations defines the Amp experience for me. I’m not embarrassed to say that I’m a big fan haha.